Realtime Systems – SMD138

Lecture 9:
Scheduling and feasibility
Burns/Wellings ch. 13 (except 13.12)

1
 Recap:
• A typical real-time OS like QNX offers
- Threads for concurrent execution
- “Parking” operations like MsgReceive(),
MsgReceivePulse(), sigwait(), ...
- Static priorities for controlling relative
importance of threads
• A system specification might contain
- A set of events to handle
- Specified reactions for each event
- A deadline for each reaction
• How map a specification onto an implementation?
REALTIME SYSTEMS – SMD138

2
 Specific problem:

How do we pick thread priorities for the purpose of meeting

deadlines?
• No generally applicable method exists...
• ... we must concentrate on programs of a rather
restricted form
• Outline:
- Start out with very restricted programs
- Reintroduce some of the complicating factors
- Discuss remaining limitations

REALTIME SYSTEMS – SMD138

3
 Priority assignment

• Under some given assumptions, there might be

several ways of assigning priorities so that deadlines
are met
• We’d like to find an optimal method, that fails only
if every other method also fails
• Two separate issues:
- Applying an optimal priority assignment method
- Determining if meeting deadlines is at all possible

REALTIME SYSTEMS – SMD138

4
 Initial restricted model
• All threads follow the event-loop pattern:
initialization();
while (1) {
parking_op();
non_blocking_code()
}
• Every parking op waits for a timer tick
• The timers are periodic with fixed period lengths
• The non-blocking code really means non-blocking;
not even mutex calls are allowed
• That is: a set of strictly periodic independent threads

REALTIME SYSTEMS – SMD138

5
 Notation
For each thread i we know
- its period Ti (given by control algorithm)
- its execution time Ci (measured, or analyzed)
- its relative deadline Di (equal to Ti for now)
Ti ( = Di )

Ci
We want to determine each thread priority Pi

REALTIME SYSTEMS – SMD138

6
 Basic rule
• Under the given assumtions, there exists an
optimal priority assignment rule that is really
simple:
Let Pi < Pj if Tj < Ti
i.e., the shorter the period, the higher the priority
• This rule is called Rate Monotonic Priority
Assignment, or RM for short
• RM is optimal, meaning that if an RM assignment
results is missed deadlines, every other priority
assignment does so too

REALTIME SYSTEMS – SMD138

7
 Example
• Given a set of periodic threads with periods
T1 = 150 ms
T2 = 120 ms
T3 = 200 ms
T4 = 75 ms
• An RM priority assignment might be
P1 = 10
P2 = 12
P3 = 9
P4 = 13

REALTIME SYSTEMS – SMD138

8
 Optimality & schedulability

• RM will produce a priority assignment that meets

deadlines, if such an assignment exists
• [ Note that for some given set of periods, there
might be other priority assignments that also give
good results ]
• However, RM doesn’t really say that the resulting
set of threads wi" meet their deadlines!
• [ This is like knowing where the shortest path from
A to B lies, but still not knowing if that path is
short enough so that B can be reached in time ]

REALTIME SYSTEMS – SMD138

9
 Utilization-based analysis
Given as set of simple periodic threads, scheduling with
priorities according to RM will succeed if
!
N
Ci
U≡ ≤ N (21/N − 1)
i=1
Ti
where N is the number of threads

That is, the sum of all CPU utilizations must be less than
a certain bound that depends on N

REALTIME SYSTEMS – SMD138

10
Utilization bounds
N Utlilization bound
1 100.0 %
2 82.8 %
3 78.0 %
4 75.7 %
5 74.3 %
10 71.8 %

Approaches 69.3% asymptotically

REALTIME SYSTEMS – SMD138

11
 Example A

Thread Period Execution time Priority Utilization

i Ti Ci Pi Ui
1 50 12 10 24%
2 40 10 11 25%
3 30 10 12 33%

The combined utilization is 82%, which is above the

bound for 3 threads (78%)
The thread set fails the utilization test

REALTIME SYSTEMS – SMD138

12
 Time-line for example A

Missed
deadline

0 10 20 30 40 50 60

REALTIME SYSTEMS – SMD138

13
 Example B

Thread Period Execution time Priority Utilization

i Ti Ci Pi Ui
1 80 32 10 40%
2 40 5 11 12.5%
3 16 4 12 25%

The combined utilization is 77.5%, which is below than

the bound for 3 threads (78%)
Hence the thread set will meet all its deadlines

REALTIME SYSTEMS – SMD138

14
 Time-line for example B

0 16 32 48 64 80 96

REALTIME SYSTEMS – SMD138

15
 Example C

Thread Period Execution time Priority Utilization

i Ti Ci Pi Ui
1 80 40 10 50%
2 40 10 11 25%
3 20 5 12 25%

The combined utilization is 100%, which is well above

the bound for 3 threads (78%)
However, the thread set still meets all its deadlines!

REALTIME SYSTEMS – SMD138

16
 Time-line for example C

0 10 20 30 40 50 60 70 80

REALTIME SYSTEMS – SMD138

17
 Characterisics

• The utilization-based test is

- sufficient (pass the test, and you’re ok)
- not necessary (fail, and you’ve still got a chance)
• Why bother with such a test?
- Because it’s so simple!

REALTIME SYSTEMS – SMD138

18
 Response time analysis
• A more elaborate test that is both sufficient and
necessary
• Idea: compute the response time for each thread, and
compare with deadline
• For highest priority thread, response time equals
execution time
• For second but highest, response time is execution
time + interference by highest priority thread
• Repeat recursively
• Interference depends on period ratio

REALTIME SYSTEMS – SMD138

19
 Response time analysis
• For each thread i, calculate its response time Ri:
! " Ri #
Ri = Ci + Cj
Tj
j∈hp(i)

where hp(i) is the set of threads with priority higher

than i, and !x" means the least integer ≥ x
• Then check if each Ri is less than the corresponding
period (=deadline) Ti
• But how calculate an Ri that depends on itself?
• Answer: start by assuming Ri = Ci , then repeat
calculation based on previous value intil it stabilizes
ProcPr – ISI217

20
 Execution times

• How can we obtain worst-case execution times?

- By measurement (but how do we know that we
are measuring the worst case?)
- By analysis (but how do we accurately model
pipelines, caches, etc?)
• How do we deal with data dependencies?
- Conservative approximations
- Use semantic information

REALTIME SYSTEMS – SMD138

21
 Example
• Assume the following code:
for (i = 1; i <= 10; i++) {
if (cond)
// basic block that costs 100 ms
else
// basic block that costs 10 ms
}
• A conservative approximation says each turn takes
100 ms; i.e., 10*100 ms = 1000 ms in total
• But now assume cond is actually i < 4. Since it can
only be true at most 3 turns, the worst case is 3*100 +
7*10 = 370 ms, which is much less than 1000 ms
REALTIME SYSTEMS – SMD138

22
 Hard vs. soft real-time

• A hard task must always meet its deadline

- Use worst-case execution time in analysis
• A soft task must be fast enough over some period
of time
- Use mean execution time in analysis
• Note: with modern architectures, the mean case
might differ from the worst case with a huge factor
(in the 1000 range)

REALTIME SYSTEMS – SMD138

23
 Loosening the restrictions
• Let deadlines be less than periods; i.e., Di ≤ Ti
- Use Deadline Monotonic priority assignment (DM)
• In addition to periodic threads, let some threads be
sporadic, with some minimum inter-arrival time Ti
- Treat these as periodic threads with period Ti, as
long as Di ≤ Ti
• Let threads block for access to common resources
- Extend Response-time analysis with Blocking
time analysis (requires Priority Inheritance)

REALTIME SYSTEMS – SMD138

24
 Aperiodic tasks
• An aperiodic task is not periodic, and lacks a
minimum inter-arrival time
• Such a task might easily consume the whole CPU
under bad circumstances (dense arrivals)
- This suggests giving aperiodic tasks the lowest
priorities
• However, under better circumstances, the relative
importance of an aperiodic task might suggest a
higher priority...
• The solution is to associate an “account” of cpu
cycles with a thread – as long as there are cycles on
the account, priority is high, otherwise it drops
• This is the meaning of a Sporadic Server in POSIX
REALTIME SYSTEMS – SMD138

25
 Deadline scheduling
For comparison, there actually exists an alternative
scheduling algorithm that
- does not apply to periodic processes only
- can handle sporadic processes via interrupts
- does not need an intermediate priority
assignment step
- has a simple utilization-based test that is
necessary and sufficient, and extremely simple:
N
∑
i =1
Ci
Ti
≤1

!
N
Ci
U≡ ≤1
i=1
Ti
This algorithm is called Earliest Deadline First (EDF)
REALTIME SYSTEMS – SMD138

26
 EDF
• EDF is a dynamic scheduling algorithm, since it
prioritizes threads according to their absolut$
deadline at run-time
- The absolute deadline of an event is the time of
the event + its relative deadline
• Like RM and fixed priorities, EDF is an optimal
scheduling algorithm
• EDF constitutes a promising alternative to RM/
DM because of its
- close relation to the terminology of real-time
specifications
- superior resource utilization
- ease of handling in the sporadic/aperiodic cases
REALTIME SYSTEMS – SMD138

27
 Drawbacks of EDF

• EDF still lacks wide-spread use because

- It exhibits random behavior under overload
- Utilization-based test becomes much more
elaborate when Di ≤ Ti
- Operating systems generally don’t support it
- Few languages allow for natural deadline
constraints
• Still, EDF continues to be intensively studied
because of its theoretical properties

REALTIME SYSTEMS – SMD138

28
 Summary
• Fixed priority scheduling applies primary to periodic
processes (sporadic ≈ periodic in the worst case)
• Aperiodic processes can be handled by the dynamic
technique of sporadic servers
• Priorities are assigned according to period (RM) or
deadline (DM) [note: RM is a special case of DM]
• Utilization-based test requires CPU utilization to be
less than 69% in the general case
• Response-time analysis is more refined, and can
handle blocking for resources
• Deadline-based scheduling (EDF) is a theoretically
appealing alternative to fixed priorities
REALTIME SYSTEMS – SMD138

29